Abstract
Resumo—Na análise de imagens hiperespectrais, uma abordagem particularmente bem-sucedida emprega bibliotecas espectrais estruturadas para lidar com a variabilidade espacial na separação espectral esparsa. No entanto, as penalidades usadas agregam um custo computacional substancial, e as soluções são sensíveis ao ruído. Neste trabalho, generalizamos a abordagem de regularização multi-escala para resolver o problema de separação espectral incorporando normas indutoras de esparsidade de grupo. Ainda, propomos um método estruturado, resistente a ruídos, para lidar com a variabilidade dos materiais e garantir a esparsidade inter-e intra-classe com reprodutibilidade e baixo custo computacional. Os experimentos ilustram a robustez e a consistência dos resultados.
Introduction
Hyperspectral Unmixing (HU) is a critical task in remote sensing, aimed at identifying the constituent materials (endmembers) and their fractional abundances within each pixel of a hyperspectral image. A significant challenge in HU is "spectral variability"—the phenomenon where a material's spectral signature varies due to factors like illumination, atmospheric conditions, and intrinsic material differences. While using spectral libraries or "bundles" can address this, it often introduces high computational costs and sensitivity to noise.
This paper introduces the Generalized Multiscale Bundle-based Unmixing Algorithm (GMBUA). GMBUA addresses these challenges by integrating multi-scale spatial processing with a graph-based consensus strategy, ensuring both robustness to noise and reproducibility of results.
Methodology
The GMBUA framework relies on two primary innovations:
- Multi-scale Spatial Regularization: Instead of processing pixels in isolation, the algorithm leverages spatial correlations by grouping similar pixels into superpixels. Unmixing is first performed at a coarse scale (the average of the superpixel), and this solution regularizes the fine-scale, pixel-by-pixel unmixing. This approach acts as a natural denoiser and improves computational efficiency.
- Graph-Based Centrality for Reproducibility: To handle the stochastic nature of spectral bundle extraction, the algorithm is executed $K$ times. A graph is constructed where nodes represent complete solutions. By calculating the similarity between these solutions (using a Linear Assignment Problem to handle permutations), the "most central" solution is identified as the representative consensus, ensuring stable results across different runs.
The algorithm uses mixed norms ($\ell_{G,r,s}$) to enforce "group sparsity," ensuring that the model picks only a few materials (inter-class sparsity) and only the most relevant signatures within each material's bundle (intra-class sparsity).
Experimental Evaluation
The researchers evaluated GMBUA against several state-of-the-art methods, including FCLSU, Collaborative, Group, Elitist, and Fractional unmixing.
- Synthetic Dataset: A $100 \times 100$ pixel image with 9 endmembers from the USGS library, tested under varying noise levels (20dB and 30dB).
- Real-World Dataset: The Houston dataset, featuring 144 spectral bands and five main classes: concrete, metal roofs, vegetation, asphalt, and colored structures.
Results and Discussion
GMBUA demonstrated superior performance, particularly in high-noise environments. Qualitative results show that GMBUA produces significantly cleaner and more consistent abundance maps compared to other methods, which often exhibit "speckled" artifacts in the presence of noise.
Quantitatively, the benefits of the multi-scale approach are evident in the Signal-to-Reconstruction Error (SRE) and Root Mean Square Error (RMSE) metrics. In the 20dB noise scenario, GMBUA achieved an SRE of 8.51 dB, vastly outperforming the Fractional method (2.87 dB) and FCLSU (3.10 dB).
While simpler methods like FCLSU may perform slightly better in noise-free conditions due to the lack of sparsity-induced bias, GMBUA's robustness makes it far more suitable for practical, real-world applications.
Conclusion
GMBUA offers a robust framework for hyperspectral unmixing by effectively combining spatial context with structured sparsity. While it requires hyperparameter tuning and multiple runs for consensus, the resulting gain in accuracy and reproducibility marks a significant advancement in handling spectral variability.
Summary updated